Elon: FSD Beta tweets

[S4WRXTTCS]

Over/under on Chuck's left turn success rate?
My bet is less than 95%.

If someone does a counter bet of $20 I'll pledge a donation of $100 to a charity/club of their choice if the initial version of 10.13 passes chucks test more than 95% of the time with a 20 test minimum.

If 10.13 fails the test they must donate the $20 to some robotics club.

It can only be one person as I'm poor.

 

[DanCar]

Most knew it wasn’t a sabbatical. Allegedly Karpathy disagreed with Elons approach of repeatedly giving unreachable FSD dates/timelines that he knew were unobtainable

Where did you get that info?

 

[n.one.one]

It's great to see Chuck Cook (@Chazman92) get recognized by Elon for his epic work on unprotected left turns.
Here was his first UPL video Nov. 22, 2020.

FSDBeta 2020.44.10.2 - Lots of unprotected left turns tested - 9:43 - Chuck Cook

 

[AlanSubie4Life]

95% is still horrible for a geofenced test that requires many 9's to achieve human parity.

I thought you were talking about the over under on what it would actually do, not what was necessary!

If someone does a counter bet of $20 I'll pledge a donation of $100 to a charity/club of their choice if the initial version of 10.13 passes chucks test more than 95% of the time with a 20 test minimum.

This would be a crazy bet, there's absolutely no way it will get anywhere near 95% success rate.


Success has to be:
1) With traffic that the car has to respond to. Can't be an empty road. (Successes where it goes with no traffic at all don't count, even though that would be an improvement from 10.12.)
2) Car has to go at the first opportunity when a decent human driver would go, when there is a nice relaxed opening. (This is the hardest thing to define but I'll know it when I see it.)
3) The car cannot cause any other car to slow down or react at any point in the maneuver (by poking out, moving forward at the wrong time), etc. Including if it stops halfway through in the median - it cannot result in any car slowing or yielding to it or anything like that.
4) The car can't do anything incorrect, like stopping in the first lanes of traffic (even if there is no one coming) to wait for traffic from the other direction.
5) Interventions including accelerator application count as failures.

I'd be VERY surprised if it reaches 30% success rate. So I'd bet less than 30%.

Someone should design a poll.

 

[Daniel in SD]

I thought you were talking about the over under on what it would actually do, not what was necessary!



This would be a crazy bet, there's absolutely no way it will get anywhere near 95% success rate.


Success has to be:
1) With traffic that the car has to respond to. Can't be an empty road. (Successes where it goes with no traffic at all don't count, even though that would be an improvement from 10.12.)
2) Car has to go at the first opportunity when a decent human driver would go, when there is a nice relaxed opening. (This is the hardest thing to define but I'll know it when I see it.)
3) The car cannot cause any other car to slow down or react at any point in the maneuver (by poking out, moving forward at the wrong time), etc. Including if it stops halfway through in the median - it cannot result in any car slowing or yielding to it or anything like that.
4) The car can't do anything incorrect, like stopping in the first lanes of traffic (even if there is no one coming) to wait for traffic from the other direction.
5) Interventions including accelerator application count as failures.

I'd be VERY surprised if it reaches 30% success rate. So I'd bet less than 30%.

Someone should design a poll.

Haha. I was assuming that it would be a clear path a good percentage of the time and wasn’t considering condition 2) failures (though if the guy behind you honks then that’s a failure).
This is a geofenced, overfit case now, I’m willing to take the over on 90% without condition #1.

 

[Cheburashka]

I wish they could just get FSD to not take those turns at all and plot the route through a path which would be easy for it, rather than pointlessly trying and failing to navigate a hard route that will never work reliably.

 

[JHCCAZ]

I wish they could just get FSD to not take those turns at all and plot the route through a path which would be easy for it, rather than pointlessly trying and failing to navigate a hard route that will never work reliably.

I would actually give them extra points for deciding to go right and then U-turn at the light. Chuck's turn is not advisable in anything but the lightest traffic, and dven then the view is dangerously blocked.

AFAIK FSD beta still doesn't execute U-turns in general - so that in itself would be an advancement, and then incorporating it into intelligent navigation planning would be outstanding.

 

[AlanSubie4Life]

Haha. I was assuming that it would be a clear path a good percentage of the time and wasn’t considering condition 2) failures (though if the guy behind you honks then that’s a failure).
This is a geofenced, overfit case now, I’m willing to take the over on 90% without condition #1.

As usual you are an eternal optimist with FSD. The endearing naïveté of an EAP HW2.5 driver…

Sure. I’ll take the under and hope for moderate traffic. Probably 60% success.

or ?

In either case, nearly certainly will be (literally) infinitely better than 10.12.

 

[Daniel in SD]

Sure. I’ll take the under and hope for moderate traffic. Probably 60% success.

or ?

No, whatever traffic Chuck tests in and including the trivial unprotected left. Disengagements, interventions, and collisions are the only failure modes.

90% is still but that’s better than the current

 

[DanCar]

I thought you were talking about the over under on what it would actually do, not what was necessary!



This would be a crazy bet, there's absolutely no way it will get anywhere near 95% success rate.


Success has to be:
1) With traffic that the car has to respond to. Can't be an empty road. (Successes where it goes with no traffic at all don't count, even though that would be an improvement from 10.12.)
2) Car has to go at the first opportunity when a decent human driver would go, when there is a nice relaxed opening. (This is the hardest thing to define but I'll know it when I see it.)
3) The car cannot cause any other car to slow down or react at any point in the maneuver (by poking out, moving forward at the wrong time), etc. Including if it stops halfway through in the median - it cannot result in any car slowing or yielding to it or anything like that.
4) The car can't do anything incorrect, like stopping in the first lanes of traffic (even if there is no one coming) to wait for traffic from the other direction.
5) Interventions including accelerator application count as failures.

I'd be VERY surprised if it reaches 30% success rate. So I'd bet less than 30%.

Someone should design a poll.

Maybe add to the list: Car can't go crazy when sun shining into camera. I noticed this for the first time a couple of days ago. True for robotaxi, not so true if a human can take over.

 

[AlanSubie4Life]

Disengagements, interventions, and collisions are the only failure modes.

I was only talking about his difficult busy left turn, not the other easy one.

I’m not sure how we’d even count if we were doing multiple turns. Equal weight of success rates which will have huge steps in % due to limited attempts? Weight proportion to how many times they were tried?

You missed the failure mode where the car causes other users to react (slow down, deviate in their lane, change lanes, etc.)

Also want to clarify that the car has to come up to the stop sign/turn without any accelerator application to position it. And each engagement/disengagement at the turn of course counts as a separate attempt & failure.

Minimum number of attempts? I assume you want more than 10!

 

[Daniel in SD]

I was only talking about his difficult busy left turn, not the other easy one.

I’m not sure how we’d even count if we were doing multiple turns. Equal weight of success rates which will have huge steps in % due to limited attempts? Weight proportion to how many times they were tried?

(number of successful turns)/(number of turns) in Chuck's video!

You missed the failure mode where the car causes other users to react (slow down, deviate in their lane, change lanes, etc.)

I don't count that, too subjective. I'm not expecting them to fix the creeping while cars are clearly visible behavior.

Also want to clarify that the car has to come up to the stop sign/turn without any accelerator application to position it. And each engagement/disengagement at the turn of course counts as a separate attempt & failure.

Yeah, no accelerator application and each disengagement is a separate failure.

Minimum number of attempts? I assume you want more than 10!

Kind of assuming he'll keep trying until there's a failure. No minimum number of attempts.

 

[AlanSubie4Life]

I don't count that, too subjective

We’ve seen this many times in prior videos and it is a huge problem with FSD. We can tighten it to obvious reactions to the Tesla. I think it should be clear.

Anyway, sure.

 

[JupiterMan]

Maybe add to the list: Car can't go crazy when sun shining into camera. I noticed this for the first time a couple of days ago. True for robotaxi, not so true if a human can take over.

actually, FSD/autopilot seems to have better view of the road ahead than I do in those conditions. to the point where I activate autopilot when driving into direct sunlight, I feel much safer.

 

[diplomat33]

actually, FSD/autopilot seems to have better view of the road ahead than I do in those conditions. to the point where I activate autopilot when driving into direct sunlight, I feel much safer.

It depends on the angle of the sun. In some instances if the sun is in your eyes but not hitting the cameras at the wrong angle, yes, it is safer to use AP. However, when the sun hits the cameras at just a certain angle, I get the "camera blinded" warning and the red hands on wheel and AP cancels on me in the middle of the drive. For maybe 15 mn, I can't use AP.

 

[DanCar]

https://twitter.com/x/status/1548975299282419712

Text version:

https://www.reddit.com/r/teslamotors/comments/w1xtgz/_/igmwmaf

 

[ThomasD]

Can it see cars at the bottom of a hill or coming up a hill while making unprotected left turns?

 

[thewishmaster]

https://twitter.com/x/status/1549138566990766082

 

[dramsey]

https://twitter.com/x/status/1548975299282419712

Text version:

https://www.reddit.com/r/teslamotors/comments/w1xtgz/_/igmwmaf

Sadly I don't see anything that addresses my biggest gripe: hilariously bad lane choices leading up to a turn, i.e. moving into the far left lane a block before a right turn.

 

[AlanSubie4Life]

Sadly I don't see anything that addresses my biggest gripe: hilariously bad lane choices leading up to a turn, i.e. moving into the far left lane a block before a right turn.

Can take the corner faster that way, and less fear of curbing a wheel.